Multi-agent reinforcement learning (MARL) excels in cooperative and competitive tasks, but most architectures are tied to fixed input-output sizes and require retraining when the number of perceptible or controllable objects changes. While structural generalization techniques mitigate this, they rely on centralized training, raising concerns about scalability and privacy. We propose ADAPT, the first framework to support structural generalization under a decentralized training and decentralized execution (DTDE) paradigm. Every agent adopts an object-centric view, encoding each observed object into a feature vector and aggregating them into a variable-length set representation. To enable each agent to infer task-level contexts from this dynamic input independently, we propose a dynamic-consistency loss that enforces spatio-temporal alignment between context representations and observed environmental dynamics. Agents then condition their policies on the inferred contexts to make locally aligned decisions. For zero-shot transfer, we propose FINE (Foresight INdex for multi-agEnt), a metric that considers Q-value overestimation and enables cross-policy comparison of long-term impact, facilitating effective policy transfer. Experiments show that ADAPT surpasses existing DTDE methods and outperforms CTDE baselines in zero-shot generalization.

As the number and complexity of constraints in constrained multi-objective optimization problems (CMOPs) increase, the performance of existing constrained multi-objective evolutionary algorithms (CMOEAs) declines significantly. A novel idea is to sequentially address each constraint based on priority, effectively reducing the complexity of CMOPs. However, in these algorithms, the constraint-handling priority is determined statically in the initial stage. This may lead to inappropriate determination of constraint-handling priority since accurately estimating the constraint landscape in the initial stage is quite challenging. Moreover, these algorithms tackle constraints separately, neglecting the potential for inter-constraint cooperation and thus compromising their efficiency in constraint handling. Thus, we propose a constrained multi-objective evolutionary algorithm based on dynamic priority and cooperative offspring generation called DPCMOEA. Firstly, the constraint-handling priority is determined dynamically by the estimated inconsistency degree (EID) between the Pareto fronts of the candidate constraints and the current population. Secondly, computational resources are automatically allocated to each constraint according to EID based constraint relationship analysis. Finally, a new offspring generation strategy based on constraint cooperation is designed to enhance the quality of new solutions. Experimental results on six CMOP test suites demonstrate that DPCMOEA outperforms six state-of-the-art algorithms.

Multi-objective Bayesian optimization (MOBO) has shown promising performance on various expensive multi-objective optimization problems (EMOPs). However, effectively modeling complex distributions of the Pareto optimal solutions is difficult with limited function evaluations. Existing Pareto set learning algorithms may exhibit considerable instability in such expensive scenarios, leading to significant deviations between the obtained solution set and the Pareto set (PS). In this paper, we propose a novel Composite Diffusion Model based Pareto Set Learning algorithm (CDM-PSL) for expensive MOBO. CDM-PSL includes both unconditional and conditional diffusion model for generating high-quality samples efficiently. Besides, we introduce a weighting method based on information entropy to balance different objectives. This method is integrated with a guiding strategy to appropriately balancing different objectives during the optimization process. Experimental results on both synthetic and real-world problems demonstrates that CDM-PSL attains superior performance compared with state-of-the-art MOBO algorithms.

Portfolio selection is a critical task in finance, involving the allocation of resources across various assets. However, current methods often struggle to maintain robust performance due to the inherent low signal-to-noise ratio in raw financial data and shifts in data distribution. We propose AlphaGAT, a novel two-stage learning approach for portfolio selection, designed to adapt to different market scenarios. Inspired by the concept of alpha factors, which transform historical market data into actionable signals, the first stage introduces an advanced model named CATimeMixer for alpha factor generation with a novel loss function to improve the effectiveness and robustness. CATimeMixer integrates TimeMixer with Conv1D (C) and cross-asset Attention (A). Specifically, Conv1D enhances TimeMixer by capturing trend and seasonal features across different scales, while cross-asset attention enables TimeMixer to extract interrelationships between different assets. The second stage applies reinforcement learning to dynamically adjust weights, integrating alpha factors into trading signals. Recognizing the varying effectiveness of alpha factors across different periods, our RL agent innovatively transforms the alpha factors into graphs and employs graph attention networks (GAT) to discern the significance of different alpha factors, enhancing policy robustness. Extensive experiments on real-world market data show that our approach outperforms state-of-the-art methods.

Collaborative search by air-ground unmanned swarm, as a pivotal and efficient approach for intelligence gathering and disaster relief, highlights the critical role of search path planning in enhancing overall performance. Addressing the inefficiency resulting from insufficient collaboration between air and ground unmanned platforms in current research, this paper delves into the fundamental characteristics and challenges of collaborative search by air-ground unmanned swarm. This paper clarifies the objectives and constraints of path planning and introduces a method for collaborative search path planning based on the Learning Wolf Pack Algorithm (LWPA). This method initially constructs an optimization model that comprehensively considers area coverage, target detection probability, and search uncertainty. It integrates Distributed Model Predictive Control (DMPC) with the Distributed Constraint Optimization Problem (DCOP) framework, forming an architecture for real-time search path planning. To overcome the limitation of existing DCOP solution methods, which tend to get stuck in local optimal solutions, the LWPA employs a Q-learning mechanism for hierarchical learning and dynamically adjusts parameters to balance local refinement and global exploration. Experimental results demonstrate that this method offers significant advantages in improving search efficiency, coverage, and target detection rates, with an average area coverage of 99.28% and uncertainty as low as 0.86%. These results fully validate its effectiveness and superiority in search tasks in complex urban environments. Furthermore, tests on dynamic adaptability and scalability further verify the potential and value of this method in practical applications.

Dynamic constrained multiobjective optimization problems (DCMOPs) require algorithms to quickly track the feasible Pareto optima under dynamic environments. The existing dynamic constrained multiobjective evolutionary algorithms (DCMOEAs) normally focus on the convergence speed, but cannot well guarantee distribution. To address this issue, a subspace sparsity driven knowledge transfer strategy based DCMOEA is developed in this article, called SSDKT. First, reference points are introduced to partition objective space into multiple subspaces. Subsequently, the feasibility of each subspace is determined by the distribution of all historical feasible optimal solutions in it, and defined as the sparsity of subspace. A predictor based on the gated recurrent unit (GRU) network is further constructed to estimate the sparsity under the future environment. Once a new environment appears, a subspace transfer strategy is designed to generate an initial population. In each feasible subspace, the GRU-based prediction method is developed and competed with Kalman filter to generate the initial solution under the new environment. Based on the predicted solution of the nearest feasible neighbor, a potential initial individual in each infeasible subspace is produced by transferring the corresponding knowledge. The experimental results on various benchmarks verify that, compared with several state-of-the-art DCMOEAs, the proposed algorithm achieves the most competitive performance in solving DCMOPs.

To solve dynamic multiobjective optimization problems (DMOPs), the optimization algorithms are required to track the movement of the Pareto set after the environmental changes effectively. Many prediction-based dynamic multiobjective evolutionary algorithms (DMOEAs) have been proposed to address this challenge by utilizing environmental information for population reinitialization. However, when environmental changes are complex, irregular, and severe, the solutions and information during the evolution process often contain noise, making it difficult for prediction-based DMOEAs to accurately predict and reinitialize the population. To address this issue, we propose a novel dynamic multiobjective evolutionary algorithm (DM-DMOEA) which uses a diffusion model-based prediction strategy. In DM-DMOEA, to improve the prediction accuracy, the diffusion model is introduced to extract the relationships of high-quality solutions and reinitialize the population, and a PS estimation method is employed to integrate both historical and new environmental information, providing a set of high-quality solutions for diffusion model training. To speed up the response time, a variational autoencoder (VAE) is used to map the decision space to a latent space, which can reduce the diffusion model size and accelerate the diffusion process. To evaluate the effectiveness of the proposed DM-DMOEA on DMOPs, comprehensive experiments are conducted on several benchmarks and a practical problem. The results show that the DM-DMOEA outperforms other four state-of-the-art DMOEAs in most cases.

Defect recognition is an important task in intelligent manufacturing. Due to the subjectivity of human annotation, the collected defect data usually contains a lot of noise and unpredictable uncertainties, which have a great negative influence on defect recognition. It is a significant challenge to discover an effective defect recognition model with satisfactory uncertainty processing ability. A natural way is to automatically search for an efficient deep model, which can be realized by neural architecture search (NAS). To achieve this, we propose an efficient fuzzy NAS framework for defect recognition, where the searched architecture can effectively handle uncertain information from the given datasets. Specifically, we first design a fuzzy search space and the related encoding strategy for fuzzy NAS. Then, we propose a comparator-based evolutionary search approach, where an online end-to-end comparator is learned to directly determine the selection of candidate architectures from the evolutionary population. The comparator works in an end-to-end way and it transforms the complex ranking problem of evaluating architectures into a simple classification task, which overcomes the rank disorder issue suffered from traditional performance predictors. A series of experimental results demonstrate that the architecture with fewer #Params (1.22 M) search by fuzzy neural architecture search framework for defect recognition method achieves higher accuracy (92.26%) compared to the state-of-the-art results (i.e., DARTS-PV) on the ELPV dataset, as well as competitive results (accuracy = 76.4%, #Params = 1.04 M) on the CODEBRIM dataset. Experimental results show the effectiveness and efficiency of our proposed method in handling uncertain problems.

Ship detection in optical satellite images has played an important role in the field of remote sensing for a long time. Many detection methods have been proposed to address the ship detection problem, and most of them mainly focus on the improvement of detection accuracy but rarely pay attention to the detection speed. In this paper, we not only consider the improvement of detection accuracy, but also try to speed up the detection process. Based on the YOLOv2 model, we propose a forward propagation acceleration-based deep neural network model (FPA-DNN) to enhance the performance of the ship detection. The FPA-DNN model is a hybrid learning model, in which the deep neural network model LSDN can effectively reduce the number of parameters and improve the detection speed with no accuracy loss, and the pruning based forward propagation acceleration algorithm can remove the redundant convolution kernels and further speed up the detection process. Experimental results on the optical remote sensing image dataset show that, compared with several state-of-the-art deep learning models, 1) the LSDN model outperforms the others on the detection accuracy and detection speed; and 2) the FPA-DNN model can further improve the detection accuracy and speed up the detection process significantly.

As one of the classical problems in the economic market, the newsvendor problem aims to make maximal profit by determining the optimal order quantity of products. However, the previous newsvendor models assume that the selling price of a product is a predefined constant and only regard the order quantity as a decision variable, which may result in an unreasonable investment decision. In this article, a new newsvendor model is first proposed, which involves of both order quantity and selling price as decision variables. In this way, the newsvendor problem is reformulated as a mixed-variable nonlinear programming problem, rather than an integer linear programming problem as in previous investigations. In order to solve the mixed-variable newsvendor problem, a histogram model-based estimation of distribution algorithm (EDA) called EDAmvnis developed, in which an adaptive-width histogram model is used to deal with the continuous variables and a learning-based histogram model is applied to deal with the discrete variables. The performance of EDAmvn was assessed on a test suite with eight representative instances generated by the orthogonal experiment design method and a real-world instance generated from real market data of Alibaba. The experimental results show that, EDAmvnoutperforms not only the state-of-the-art mixed-variable evolutionary algorithms, but also a commercial software, i.e., Lingo.

Knowledge withholding, which is defined as the likelihood that an individual devotes less than full effort to knowledge contribution, can be regarded as an emerging social deviance behavior for knowledge practice in online knowledge spaces. However, prior studies placed a great emphasis on proactive knowledge behaviors, such as knowledge sharing and contribution, but failed to consider the uniqueness of knowledge withholding. To capture the social‐deviant nature of knowledge withholding and to better understand how people deal with counterproductive knowledge behaviors, this study develops a research model based on the secondary control perspective. Empirical analyses were conducted using the data collected from an online knowledge space. The results indicate that both predictive control and vicarious control exert a positive influence on knowledge withholding. This study also incorporates knowledge‐withholding acceptability as a moderating variable of secondary control strategies. In particular, knowledge‐withholding acceptability enhances the impact of predictive control, whereas it weakens the effect of vicarious control on knowledge withholding. This study concludes with a discussion of the key findings, and the implications for both research and practice.

Feature selection is an effective technique for dimensionality reduction to get the most useful information from huge raw data. Many spectral feature selection algorithms have been proposed to address the unsupervised feature selection problem, but most of them fail to pay attention to the noises induced during the feature selection process. In this paper, we not only consider the feature structural learning, but also try to avoid these noises induced during the feature selection process. We utilize the feature structural learning to select the discriminant features and use the robust methods to make selected features more reliable. Furthermore, we propose a new unsupervised feature selection algorithm, namely Robust Sparse Subspace Learning Feature Selection(RSS). And we employ a coordinate descendent algorithm to solve the RSS formulation. Experiments are conducted on several popular datasets to validate the effectiveness of our proposed algorithm and results show that this RSS algorithm achieves better results than traditional feature selection algorithms in most cases.

Researchers have widely investigated the fractal property of complex networks, in which the fractal dimension is normally evaluated by box-covering method. The crux of box-covering method is to find the solution with minimum number of boxes to tile the whole network. Here, we introduce a particle swarm optimization box-covering (PSOBC) algorithm based on discrete framework. Compared with our former algorithm, the new algorithm can map the search space from continuous to discrete one, and reduce the time complexity significantly. Moreover, because many real-world networks are weighted networks, we also extend our approach to weighted networks, which makes the algorithm more useful on practice. Experiment results on multiple benchmark networks compared with state-of-the-art algorithms show that this PSOBC algorithm is effective and promising on various network structures.

The fractality property are discovered on complex networks through renormalizaiton procedure, which is implemented by box-covering method. The unsolved problem of box-covering method is finding the minimum number of boxes to cover the whole network. Here, we introduce a differential evolution box-covering algorithm based on greedy graph coloring approach. We apply our algorithm on some benchmark networks with different structures, such as a E.coli metabolic network, which has low clustering coefficient and high modularity; a Clustered scale-free network, which has high clustering coefficient and low modularity; and some community networks (the Politics books network, the Dolphins network, and the American football games network), which have high clustering coefficient. Experimental results show that our algorithm can get better results than state of art algorithms in most cases, especially has significant improvement in clustered community networks.

Stock price volatility prediction is regarded as one of the most attractive and meaningful research issues in financial market. Some existing researches have pointed out that both the prediction accuracy and the prediction speed are the most important facts in the process of stock prediction. In this paper, we focus on the problem of how to design a methodology which can improve prediction accuracy as well speed up prediction process, and propose a multi-kernel learning based extreme learning machine (MKL-ELM) model to enhance the prediction performance. ELM is a fast learning model and has been successfully applied in many research fields. Based on ELM, this MKL-ELM has the benefits of both multiple kernel learning and ELM, which can well balanced the requirements of both prediction accuracy and prediction speed. To validate the performance of the proposed MKL-ELM, we take experiments on HKEx 2001 stock market datasets. The market historical price and the market news are implemented in our MKL-ELM. We Compare our proposed MKL-ELM with Back-Propagation Neural Network(BP-NN), Support Vector Machine(SVM), Basic ELM and K-ELM. Experimental results show that, 1) MKL-ELM, K-ELM and SVM get higher prediction accuracy than BP-NN and B-ELM; 2) Both MKL-ELM and K-ELM can achieve faster prediction speed than SVM in most cases; 3) MKL-ELM has higher prediction accuracy in some cases than K-ELM and SVM.

This paper proposes a new approach named SGMIEC in the field of Estimation of Distribution Algorithm (EDA). While the current EDAs require much time in the statistical learning process as the relationships among the variables are too complicated, the Selfish Gene Theory (SG) is deployed in this approach and a Mutual Information and Entropy based Cluster (MIEC) model with an incremental learning and resample scheme is also set to optimize the probability distribution of the virtual population. Experimental results on several benchmark problems demonstrate that, compared with BMDA and COMIT , SGMIEC often performs better in convergent reliability, convergent velocity and convergent process.

Evolutionary Hardware (EHW) is a new focus in recent research work. The new method of design hardware is combined evolution algorithm with programmable logic device. Optimization digital circuit is a main domain of EHW. The algebra way and Karnaugh map way are the traditionary methods, but they will meet trouble with the large scale ones to get optimization structure of circuit. This paper proposes a new method (GEP) to optimize the complex digital circuit and designs a new function fitness. The experiments demonstrate the GEP is not only fast convergence but also optimization large circuit. It conquers the slow convergence even not convergence of the traditionary method. The GEP algorithm is simpler and more efficient than the traditional ones.

The Analog circuits are very important in many high-speed applications such as communications. Since the size of analog circuit is becoming larger and more complex, the design is becoming more and more difficult. This paper proposes a new circuit representation method based on a two-layer evolutionary scheme with Genetic Programming (TLGP), which uses a divide-and-conquer approach to evolve the analog circuits. This representation has the desirable property which is more helpful to generate expectant circuit graphs. And it is capable of generating various kinds of circuits by evolving the circuits with dynamical size, circuit topology, and component values. The experimental results on the designs of the voltage amplifier and the low-pass filter show that this method is efficient.

The non-existence of triangular arbitrage in an efficient foreign exchange markets is widely believed. In this paper, we deploy a forecasting model to predict foreign exchange rates and apply the triangular arbitrage model to evaluate the possibility of an arbitrage opportunity. Surprisingly, we substantiate the existence of triangular arbitrage opportunities in the exchange rate forecasting market even with transaction costs. This also implies the inefficiency of the market and potential market threats of profit-seeking investors. In our experiments, neural network based model with back-propagation (BP-NN) is used for exchange rate forecasting.